VisionOPS

About the Project

VisionOPS is a no-code AI computer vision training platform that puts the full power of deep learning into the hands of non-technical users. From raw images and video footage to a production-ready model — everything happens inside a single browser-based interface. Teams can build, train, evaluate, and download object detection and image classification models without installing a single dependency or writing a single line of code.

Dataset Management

• Upload images individually or in bulk — supports JPG, PNG, BMP, and TIFF formats
• Automatic frame extraction from uploaded video files — select FPS rate to control dataset density instantly
• In-browser image editor for cropping, rotating, brightness/contrast adjustments, and resizing before annotation
• Dataset pre-analysis engine: automatically detects class imbalance, duplicate images, low-resolution samples, mislabelled bounding boxes, empty annotation files, and truncated images — with one-click auto-fix for correctable issues
• Improvement recommendations engine: suggests augmentation strategies, minimum sample counts per class, and labelling consistency fixes based on dataset statistics

Annotation

• Built-in bounding box annotation tool with class selector, zoom, and pan — no external labelling tool needed
• Polygon and segmentation mask annotation support for instance segmentation tasks
• Keyboard shortcut-driven labelling workflow for high-speed annotation
• Annotation export in YOLO TXT, COCO JSON, Pascal VOC XML formats
• Auto-annotation assistance: AI pre-labels images using an already-trained model so annotators only review and correct

Model Training

• Full YOLOv4 (Darknet), YOLOv8, YOLOv10, YOLOv11, YOLOv12, YOLOv26 support for object detection and instance segmentation
• Classification model training with ResNet-18/34/50/101, MobileNetV2/V3, EfficientNet-B0 to B7, and VGG-16/19
• Configurable hyperparameters: epochs, batch size, learning rate, image size, augmentation pipeline, optimizer, and early stopping
• One-click training start — model selection, dataset split (train/val/test), and config generation handled automatically
• Training controls: start, stop, pause, and resume — training state is persisted so paused runs can continue without data loss
• Transfer learning from pre-trained COCO weights or custom checkpoint uploads to accelerate convergence on small datasets

Real-time Training Monitoring

• Live training dashboard with real-time charts: loss curves (box loss, object loss, class loss), mAP@0.5, mAP@0.5:0.95, precision, recall, and F1 score — updated every epoch
• Per-class metrics table with individual AP, precision, and recall breakdowns
• Confusion matrix visualisation with interactive class-level drill-down
• GPU utilisation, memory usage, and estimated time-to-completion displayed live
• Training log stream — raw stdout from the training process streamed directly to the browser console panel

Model Comparison & Analysis

• Side-by-side multi-model comparison dashboard — select any two or more trained models and compare mAP, precision, recall, F1, inference speed, and model size
• Inference test panel — upload a test image and run it against any trained model directly in the browser to see predictions with confidence scores and bounding boxes
• Model version history: every training run is saved with its configuration, metrics, and weights for full reproducibility
• One-click model download — export trained weights in PyTorch (.pt), ONNX, TensorFlow SavedModel, or TFLite format for edge deployment

Technical Architecture

The backend is built on Python with Express-style routing, handling training job orchestration, dataset storage on Google Cloud Storage (GCS), and secure API access via JWT authentication. Training jobs run as isolated Python processes, communicating progress back to the frontend via server-sent events. The dataset and model metadata are stored in MongoDB, while raw image files and model weights are stored in GCS buckets for scalability. The frontend is a React SPA built with Craco, using Axios for all API communication, and delivers a responsive, data-dense UI optimised for both data scientists and non-technical users.

Dataset Versioning & Analysis

• Create multiple independent datasets per project — organise by task type, capture session, or experiment
• Fork any dataset into a new version: add fresh images or re-annotate existing ones without touching the original, preserving full lineage
• Dataset diff view — visual side-by-side comparison of two dataset versions showing exactly which images were added, removed, or re-annotated
• Cross-version training result comparison: train on dataset v1 and v2 side by side and compare how adding 200 new images to a class shifted mAP, precision, recall, and F1
• Dataset statistics dashboard: image count per class, annotation density heatmap, bounding box size distribution, and train/val/test split visualisation
• Merge two datasets into a unified new dataset with automatic deduplication and conflict resolution for overlapping annotations
• Snapshot locking — freeze a dataset version so it cannot be modified, ensuring full reproducibility of any historical training run

Model Management

• Central model registry: every trained model is catalogued with its source dataset version, architecture, hyperparameters, training duration, hardware used, and final metrics
• Model versioning — train the same architecture multiple times across different dataset versions or config changes; all versions are stored, searchable, and independently accessible
• Tag models as Production, Staging, or Archived — instantly identify which version is deployed and which is under evaluation
• Registry search and filter by architecture type, mAP threshold, dataset version, training date range, or custom tags
• Direct model-to-model lineage graph: see which dataset version and base checkpoint each model was derived from
• Bulk actions: delete, archive, or batch-export selected model versions in a single operation
• Download any registered model version at any time in PyTorch (.pt), ONNX, TFLite, or TensorFlow SavedModel format

Industrial Integration & Enterprise Deployment

• REST API endpoints for every platform action — trigger training, query model status, fetch results, and download weights programmatically from any external system
• Embed trained models directly into existing company AI pipelines via the VisionOPS API without any code changes on the platform side
• ONNX and TFLite export ensures models are deployment-ready for edge devices, embedded systems, industrial cameras, conveyor-line inspection units, and NVIDIA Jetson boards
• Webhook notifications — receive real-time alerts in external systems (ERP, MES, SCADA, Slack, Teams) when a training run completes or a dataset version is published
• Role-based access control with team workspaces — different departments or clients can each manage their own datasets and models within a shared organisation account

Sector Coverage

VisionOPS is designed to serve any industry that benefits from visual AI — without requiring in-house data science expertise:

• Manufacturing & Quality Control — surface defect detection, dimensional inspection, assembly verification on production lines
• Agriculture — crop disease detection, weed classification, fruit counting and grading from drone or field camera footage
• Retail & Logistics — shelf stock monitoring, package damage detection, barcode and label reading automation
• Healthcare & Life Sciences — medical image pre-screening, pathology slide classification, equipment monitoring
• Security & Surveillance — person and vehicle detection, restricted area monitoring, PPE compliance verification
• Construction & Infrastructure — crack detection on structures, site safety monitoring, progress tracking from drone imagery
• Energy & Utilities — solar panel fault detection, powerline inspection, meter reading automation

Any team — regardless of technical background — can build, iterate, and deploy production-grade computer vision models tailored to their industry, and integrate the results directly into their existing software ecosystem through the VisionOPS API.

Mobile & Responsive Design

VisionOPS is fully responsive and designed to work across desktop, tablet, and mobile screens. The annotation workspace, training dashboards, real-time metric charts, dataset management panels, and model registry are all optimised for touch-friendly interaction and adaptive layouts — so team members can review training progress, approve dataset changes, and inspect model results from any device, anywhere.